Research On End-to-End Speech Recognition Based On GRU And Self-Attention Mechanism

With the improvement of computing power and the accumulation of big data corpus,speech recognition technology develops rapidly,the accuracy is greatly improved,and the application scenarios are more and more extensive.As a bridge between human and intelligent hardware devices,speech recognition has attracted more and more attention.This paper proposes an end-to-end speech recognition method,which combines the self attention mechanism with the traditional speech recognition model.It applies the end-to-end framework of link timing classification to the speech recognition task,and uses the self attention mechanism as the language model of the end-to-end speech recognition,so that the system can learn the characteristics of signals more comprehensively,and then better complete the task of Chinese speech recognition.In addition,due to the problems of high computational complexity and long training time in the long-term memory network applied to speech recognition,this paper uses the gate control cycle unit network instead of LSTM to reduce the computational cost and accelerate the training speed.According to the control experiment,under the same experimental conditions,the average training time of Gru network is 17.59% less than that of LSTM network.In this paper,GMM-HMM model baseline experiment is used to verify that the end-to-end model based on LSTM and GRU neural network has better performance in accuracy.In order to improve the accuracy of baseline experiment,bottleneck feature is used instead of MFCC feature in feature extraction.Experiments show that the bottleneck feature has stronger discrimination,which can improve the system robustness and recognition effect.